CLAUDE.md or memory server — pick the right tool
A six-question decision tree that tells you whether a markdown file is enough or whether you need a structured memory server. Most projects do not need a server.
CLAUDE.md or memory server — pick the right tool
The honest answer most memory vendors will not give you: most projects do not need a memory server. A well-maintained CLAUDE.md or AGENTS.md is enough. This recipe is the decision tree that tells you when that flips, and what to do at each level.
Step 1: Understand the three memory layers
Before deciding, know what you are choosing between. There are three layers in the market that all get sold as "memory" but are not the same thing.
- Static notes — CLAUDE.md, AGENTS.md, .cursorrules. Markdown files you edit by hand. Always loaded into context. Zero algorithm.
- Accumulating notes — Claude Code Auto-Memory (default-on since March 2026), ChatGPT Memory. The AI writes notes itself, consolidated nightly. Local or vendor-bound.
- Structured memory — Mem0, Zep, Letta, StudioMeyer Memory, etc. Knowledge graph, semantic search, confidence decay, bi-temporal model. Cloud or self-host.
Each layer solves a different problem. The decision is not "memory yes or no" — it is "which layer."
Step 2: Check duration — is the project longer than two weeks?
If no: you do not need any memory layer. Just a chat history. Move on.
If yes: continue to Step 3. This is the absolute minimum bar — anything shorter than two weeks does not accumulate enough decisions to make memory worth the setup time.
Step 3: Check repetition — do you re-explain conventions every session?
"Use TypeScript strict mode." "Never run prisma db push without a backup." "Keep responses under three paragraphs."
If no: chat history is still enough.
If yes: write a CLAUDE.md (Recipe 1.1). Stop here. This solves 80 percent of the problem for 80 percent of users. CLAUDE.md is the cheapest, most-portable, lowest-maintenance memory layer in existence — a markdown file in your repo.
Step 4: Check scale — six months and multiple days per week?
If no: CLAUDE.md is still enough. Adding a memory server creates more maintenance than benefit at this scale.
If yes: a structured memory server starts to pay off — but only if Step 5 also says yes. Single-person, single-tool projects almost never need a server.
Step 5: Check sharing — multi-tool or multi-person?
The killer use case for a memory server is shared state. You + Claude Code + Cursor + Codex on the same project. Your team plus three AI clients on the same domain knowledge.
If only you in a single tool: stick with CLAUDE.md plus Auto-Memory (Recipe 4.7 covers Auto-Memory specifically). Going further is overkill.
If multi-tool or multi-person: a memory server with MCP is the right answer. Move to Recipe 4.1 if you have not already.
Step 6: Check compliance — HIPAA, SOC 2, EU hosting?
Most memory servers do not check all the boxes by default. If your industry requires this, you need to filter the server choice on these specifically.
Examples that do most of the boxes today:
- Zep: SOC 2 Type II, HIPAA BAA. US-default.
- StudioMeyer Memory: EU hosting Frankfurt, GDPR-ready, DPA. SOC 2 not yet.
- Letta self-hosted: full control, you own the compliance story end-to-end.
If compliance is not a thing for your project: skip this question and continue.
Step 7: Check longevity — do you want this memory in five years?
If the vendor disappears or pivots, can you export your data?
- CLAUDE.md, AGENTS.md: yes, it is a file in your repo.
- Self-hosted Letta or local-memory-mcp: yes, your machine.
- StudioMeyer Memory, Mem0, Zep: yes via export endpoints, but check the format — flat JSON is portable, vendor-specific blobs are not.
- ChatGPT Memory, Anthropic Claude Memory: no full export today.
If long-term ownership matters, lean toward formats with clean export. If you are okay with vendor lock for the convenience, that is a valid trade — just make it consciously.
The summary table
| If you said yes to... | Do this | |---|---| | Step 2 only | Nothing yet. Use chat. | | Step 2 + Step 3 | Write CLAUDE.md. Recipe 1.1. | | Step 2 + Step 3 + Step 4 | Add Auto-Memory. Recipe 4.7. | | Step 2 + Step 3 + Step 4 + Step 5 | Memory server. Recipe 4.1. | | ...plus Step 6 | Filter server choice on compliance. Today: Zep or self-host. | | ...plus Step 7 | Confirm export format works for you before committing. |
What this recipe does not say
It does not say "always use the most advanced layer." Most projects regret that decision because more memory means more maintenance, more potential for stale facts, more sycophancy amplification (Recipe 4.9), more context pollution.
It does not say "memory servers are a waste." For projects that hit the criteria, they pay off enormously — saved hours per week, less context-token spend, reusable team knowledge.
The recipe says: pick the smallest layer that solves your problem. Move up only when you outgrow it.
What's next
If your decision was CLAUDE.md only, you are done with this phase — go to Phase 5 (daily use patterns).
If your decision included memory server, Recipe 4.7 covers Auto-Memory's relationship to your own server, Recipe 4.8 compares the major server options head-to-head, and Recipe 4.9 covers the risks no vendor talks about.